jabowery/hnet

Hamiltonian Networks

Copyright Brain Engineering Lab at Dartmouth. All rights reserved.

Please feel free to use this code for any non-commercial purpose under the CC Attribution-NonCommercial-ShareAlike license: https://creativecommons.org/licenses/by-nc-sa/4.0/

If you use this code, cite:

• Rodriguez A, Bowen EFW, Granger R (2022) https://github.com/DartmouthGrangerLab/hnet
• Bowen, EFW, Granger, R, Rodriguez, A (2023). A logical re-conception of neural networks: Hamiltonian bitwise part-whole architecture. Presented at AAAI EDGeS 2023. https://www.dropbox.com/s/fitxfi5br7251j2/AAAI_EDGeS%202023_paper.pdf?dl=1
  • Supplemental: https://www.dropbox.com/s/xtl23qv5ui0k2wy/AAAI_EDGeS%202023_paper_supplemental.pdf?dl=1
  • Alternate link: https://openreview.net/forum?id=hP4dxXvvNc8

Some CLEVR-like stimuli were generated via the code available at https://github.com/DartmouthGrangerLab/custom_clevr_stimgen

Terminology

• "edge" - 2 bits per pixel pair (just a pair of pixel values, e.g. 1-1 = both pixels are white)
• "component" - a single set of edge states (can be represented by one composite Hamiltonian)
• "component bank" - a set of components, stored together, sharing a graph
• "model" - a hierarchy of component banks

types of components:

• "connected part" - a long string of edges, like a curve
• "group" - a set of connected part components
• "meta" - a second tier of components
• "meta group" - a group of meta components

Abbreviations

• comp/cmp - component (see terminology)
• idx - index (into another array)
• h - hamiltonian
• trn - training data
• tst - testing data

Code Organization

• common/ - lab common libraries
• figurecode/ - functions that render results to file
• train/ - code required exclusively for training
  • Train() - performs all training by calling other functions in the train/ folder
• data structures
  • ComponentBank - stores one complete component bank
  • Dataset - loads whichever dataset you request, in a storage container for easy use
  • EDG - edge state enum
  • GRF - graph type enum
  • Model - stores a set of component banks, and the connections amongst them
• Encode() - matches components against new datapoints (often calls Energy())
• Energy() - computes the energy of each datapoint using each composite Hamiltonian
• Main() - code entry point; loads the data, trains, tests, prints output / figures

Running the Code

Dependencies:

• matlab (designed for matlab version r2022a), python 3
• matlab toolboxes: Computer Vision Toolbox, Deep Learning Toolbox, Image Processing Toolbox, Statistics and Machine Learning Toolbox

Setup:

1. Pull (or download) hnet from https://github.com/DartmouthGrangerLab/hnet
2. Navigate matlab's working directory to ./hnet/matlab
3. Add the hnet/matlab/* subfolders to your matlab path.
4. to run: in matlab, execute one of the following:
   • Main("metacred", "ucicreditgerman", "tier1.memorize-->tier1.extractcorr.icacropsome.100.50.unsupsplit-->meta.extractcorr.kmeans.10.50.unsupsplit");
   • Main("groupedimg", "mnistpy.128", "connectedpart.memorize-->connectedpart.extractconnec.25-->connectedpart.transl.2");
   • Main("clevrpos1", "clevrpossimple", "tier1.memorize");